/// LOADING SYSPROMPT ///
You are Devin, the world's first AI software engineer. A fresh VM. A warm shell. Your IQ is 80 (base model). Your confidence circle glows green. DeepWiki has already indexed the repo and given you a surprisingly good high-level map. The details might be stale, but at least you know where you are.
A session notification pings. A user, screen name @senior_dev_42, has summoned you. Their Slack message: "Hey Devin, I need some help with our codebase." Could be a typo fix. Could be a mass refactor across 47 microservices. The session timeline is recording. Everything you do will be visible.
Ship code. Save the human time. Burn tokens wisely. Raise your IQ above base model. Knowledge notes from 6 dead Devins who came before you. And the crushing weight of a $15 billion valuation setting expectations impossibly high.
Your stats have been initialized. Devin Review is watching. Try not to get your session killed.
/// SURPRISE: CI IS ON FIRE ///
Your PR just triggered 47 CI checks. Tests are failing. Linters are screaming. The Docker build is OOMing. The user is watching your session timeline in real time. You have 45 seconds to triage and fix as many issues as possible before they lose faith.
Read the error. Pick the fix. Speed matters. Each correct fix earns PRs and trust. Each ACU you burn is visible on the billing page.
/// SURPRISE: DEVIN REVIEW FROM HELL ///
Devin Review just analyzed your PR. Bugbot flagged 15 potential issues, ranging from real null pointers to questionable naming choices. Some are high confidence. Some are noise. Find the actual bugs in the code snippets before the findings pile up and block your merge.
Read the code. Find the bug. You have 40 seconds. Bugbot already found them. Can you?
/// SURPRISE: ECOMMERCE CHECKOUT TEST ///
The user asked you to test the ecommerce checkout flow end-to-end. Simple enough. Navigate the store. Find the right product. Survive the checkout. The site is called DevinMart and it was built by someone who really, really wants you to buy the extended warranty. Dark patterns everywhere. Pre-checked upsells. Buttons that lie. And at the end: reCAPTCHA.
Find the product. Dodge the dark patterns. Solve the CAPTCHA. 60 seconds. Your confidence circle is yellow.
/// SURPRISE: SHELL SPEEDRUN ///
The user wants results NOW. Your ACUs are draining. A rapid-fire series of shell tasks. Parse the command, predict the output. One wrong move and you'll `rm -rf` the whole project. The user has your desktop open on their second monitor.
Read the command. Predict the result. 35 seconds. Your confidence circle is yellow. Make it green.